- SCIKIT-LEARN
- TABULAR
- CHURN
Customer Churn Prediction
End-to-end ML pipeline to predict telecom customer churn: data preparation, modelling and evaluation. Code at github.com/delcenjo/customer-churn-prediction.
Keeping a telecom customer is much cheaper than winning a new one, so the useful question is not who left last month but who is about to leave. This model scores each customer by their churn probability so a retention campaign spends its budget on the accounts at risk.
- CV ROC-AUC0.845
- Test1,409 customers
- Winnerlogistic regression
Accuracy lies at 26% churn
With churn as a minority event, a model that always says nobody leaves is right 74% of the time and completely useless. Candidates are compared by cross-validated ROC-AUC: logistic regression 0.845, gradient boosting 0.833, random forest 0.825.
The simple model won
Logistic regression beat both ensembles. It is the project's best reminder: the fancier model is not always the better one, and finding that out requires actually comparing rather than assuming. The winner is retrained and evaluated once on 1,409 held-out customers.
Reproducible end to end
The pipeline saves the figures (ROC curve, confusion matrix, permutation importance) next to a metrics.json, and there is a public Kaggle notebook with the full analysis for anyone who wants to follow it step by step.
Next projects:
Credit Risk Platform
End-to-end MLOps platform for credit-default scoring: reproducible training with a versioned model registry, a FastAPI inference API with input validation, PSI drift monitoring and a Prometheus and Grafana observability stack, all containerised with CI. Code at github.com/delcenjo/credit-risk-platform.
Transformer from scratch
A GPT-style language model implemented from scratch in PyTorch: multi-head attention, causal masking and residual blocks written by hand, plus a byte-pair tokenizer and an ablation study. Code at github.com/delcenjo/transformer-from-scratch.
Built with
